KR20180046062A - Method for speech endpoint detection using normalizaion and apparatus thereof - Google Patents

Abstract

The present invention relates to a method for determining a voice section in an input voice signal. More specifically, the present invention relates to a method for determining a voice section using normalization and a device for determining a voice section therefor, which detect a voice endpoint by using a deep neural network (DNN) and perform normalization in a window unit in consideration of an input layer of the DNN in the normalization process before determining a voice section by using the voice endpoint, thereby enabling a more rapid process. Also, the normalization is performed after a masking process performs, thereby more accurately detecting the voice section even in various noise environments. The method for determining a voice section using normalization comprises the following steps of: extracting a voice feature vector from the input voice signal; performing the normalization on a series of voice feature vector according to the input layer of the DNN; and setting the normalized voice feature vector as an input of the DNN, and determining the voice section in the voice signal.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention [0001] The present invention relates to a method for determining a speech interval using normalization,

The present invention relates to a method for determining a speech interval in an input speech signal, and more particularly, to a speech recognition method capable of real-time processing when determining a speech interval in a speech signal and at the same time, To a method for determining a speech interval using the normalization and a speech interval determination apparatus therefor.

The contents described in this section merely provide background information on the present embodiment and do not constitute the prior art.

In order to improve the speech recognition performance, it is necessary to accurately obtain the interval of the speech applied with the noise. Accordingly, a technology for detecting an end point of a voice signal from a start point of a start of a voice to an end point of ending the voice, and thereby accurately determining a voice interval and a non-voice interval is gradually increasing in importance.

In particular, the feature normalization process is performed to more accurately detect the speech interval in consideration of various noise environments.

The feature normalization process is a GMVN (Global Mean and Variance Normalization) method in which normalization is performed by applying a previously calculated mean and variance to training data corresponding to input speech signals, And Utterance Mean and Variance Normalization (UMVN) for performing normalization by calculating variance.

The GMVN method is advantageous in that real-time processing can be performed by performing normalization by applying a previously calculated average and variance corresponding to a voice signal. However, when the noise environment in the voice signal is different from the noise environment in the training data, There is a problem in that it drops.

In the case of the UMVN scheme, the average and variance are calculated and normalized for each input speech signal. However, the performance of the UMVN scheme is better than that of the GMVN scheme. However, There is a problem that it is impossible to perform real-time processing in view of the possibility of distributed calculation.

Therefore, it is necessary to develop a feature normalization technique that can perform real time processing and show more robust performance in various noise environments.

Korean Patent Laid-Open No. 10-1999-001828, January 15, 1999 (Name: Apparatus and method for extracting speech features by dynamic region normalization of spectrum)

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and it is an object of the present invention to provide a speech interval determination method using normalization capable of performing real time processing in determining a speech interval in a speech signal, And an object of the present invention is to provide a voice interval determination apparatus therefor.

Particularly, in the present invention, normalization is performed on a per-window basis in consideration of the input layer of the depth-area neural network in the normalization process before detecting a voice end point using a deep neural network (Deep Neural Network) And it is an object of the present invention to provide a voice interval determination method using normalization capable of detecting a voice interval more accurately even in various noisy environments and a voice interval determination apparatus therefor.

However, the object of the present invention is not limited to the above-mentioned objects, and other objects not mentioned can be clearly understood from the following description.

According to an aspect of the present invention, there is provided a method for determining a speech interval using normalization, the method comprising: dividing a speech signal into frames having a predetermined length; An apparatus for determining a speech interval, the apparatus comprising: extracting a speech feature vector from an input speech signal; Performing a normalization on a series of speech feature vectors according to an input layer of a deep neural network (DNN); And determining the speech interval in the speech signal by setting the normalized speech feature vector as an input of the neural network model.

Estimating a part of the input speech signal as a noise section before extracting the speech feature vector; And masking the noise signal corresponding to the estimated noise period to the entire speech signal. The extracting of the speech feature vector may include extracting a speech feature vector from the masked speech signal .

In addition, the noise signal may be a learning noise signal applied at the time of constructing learning data for determining the speech interval.

At this time, the masking process may be performed after adjusting the energy level of the learning noise signal to match the energy level of the noise signal.

The performing of the normalization may include setting a splice window within a certain range based on an arbitrary frame in the speech signal, Calculating an average and a standard deviation corresponding to the splice window by using a speech feature vector extracted for each frequency band of the splice window; And normalizing the splice window using the calculated average and standard deviation.

At this time, the splice window may be set corresponding to the size of the input layer of the depth neural network model.

Further, the present invention can provide a computer-readable recording medium on which a program for executing the above-described method is recorded.

According to an aspect of the present invention, there is provided an apparatus for dividing a speech signal according to an exemplary embodiment of the present invention into frames each having a predetermined length, A feature vector extractor for extracting a speech feature vector from an input speech signal; A normalization unit for performing normalization on a series of speech feature vectors extracted through the feature vector extraction unit according to an input layer of a deep neural network (DNN); And a speech interval determiner for determining the speech interval in the speech signal by setting the normalized speech feature vector as an input to the depth neural network model.

The apparatus may further include a masking unit that estimates a part of the input speech signal as a noise section and adds a noise signal corresponding to the estimated noise section to the entire speech signal.

According to the speech interval determination method and the speech interval determination apparatus using the normalization of the present invention, it is possible to perform more robust performance than in the conventional normalization method in a variety of noise environments, and real-time processing is possible.

Accordingly, the present invention can more accurately detect a voice section and a non-voice section in a voice signal, and more quickly and accurately detect a voice section and perform voice recognition than in various noise environments.

Accordingly, the present invention can improve the efficiency of the speech signal processing field such as speech recognition, speaker recognition, and sound quality improvement.

In addition, various effects other than the above-described effects can be directly or implicitly disclosed in the detailed description according to the embodiment of the present invention to be described later.

1 is a block diagram illustrating a speech recognition system according to an embodiment of the present invention.
2 is an exemplary diagram for explaining a voice section according to an embodiment of the present invention.
3 is a block diagram illustrating a speech interval determination apparatus according to an embodiment of the present invention.
4 to 6 are exemplary diagrams for explaining a voice interval determination method using normalization according to an embodiment of the present invention.
7 is a flowchart illustrating a method of determining a voice interval using normalization according to an embodiment of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS For a more complete understanding of the nature and advantages of the present invention, reference should be made to the following detailed description taken in conjunction with the accompanying drawings, in which:

In the following description and the accompanying drawings, detailed description of well-known functions or constructions that may obscure the subject matter of the present invention will be omitted. It should be noted that the same constituent elements are denoted by the same reference numerals as possible throughout the drawings.

The terms and words used in the following description and drawings are not to be construed in an ordinary sense or a dictionary, and the inventor can properly define his or her invention as a concept of a term to be described in the best way It should be construed as meaning and concept consistent with the technical idea of the present invention. Therefore, the embodiments described in the present specification and the configurations shown in the drawings are merely the most preferred embodiments of the present invention, and not all of the technical ideas of the present invention are described. Therefore, It is to be understood that equivalents and modifications are possible.

Also, terms including ordinal numbers such as first, second, etc. are used to describe various elements, and are used only for the purpose of distinguishing one element from another, Not used. For example, without departing from the scope of the present invention, the second component may be referred to as a first component, and similarly, the first component may also be referred to as a second component.

In addition, when referring to an element as being "connected" or "connected" to another element, it means that it can be connected or connected logically or physically. In other words, it is to be understood that although an element may be directly connected or connected to another element, there may be other elements in between, or indirectly connected or connected.

Also, the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular expressions include plural expressions unless the context clearly dictates otherwise. It is also to be understood that the terms such as " comprising "or" having ", as used herein, are intended to specify the presence of stated features, integers, It should be understood that the foregoing does not preclude the presence or addition of other features, numbers, steps, operations, elements, parts, or combinations thereof.

Hereinafter, a voice interval determination method using normalization and a voice interval determination apparatus according to an embodiment of the present invention will be described.

FIG. 1 is a block diagram for explaining a speech recognition system according to an embodiment of the present invention, and FIG. 2 is an exemplary view for explaining a speech section according to an embodiment of the present invention.

Referring to FIG. 1, a speech recognition system 500 to which an embodiment of the present invention is applied may include a speech interval determination apparatus 100, a feature extraction apparatus 200, and a speech recognition apparatus 300. The voice segment determination apparatus 100, the feature extraction apparatus 200, and the voice recognition apparatus 300 may exist as one module that constitutes one voice recognition system 500 independent of each other, Or may be in the form of separate individual devices that can be logically connected. In addition, the voice signal inputted to the voice recognition system 500 of the present invention may be an analog voice signal, and in this case, further includes an A / D (Analog to Digital) conversion device for converting the voice signal into a digital voice signal . In addition, the speech interval determination apparatus 100 of the present invention may process the process according to an implementation method.

Each of the components configuring the speech recognition system 500 will be described. First, the speech interval determination apparatus 100 of the present invention detects an end point of an input speech signal and determines a speech interval. For example, as shown in FIG. 2, the speech interval determination apparatus 100 of the present invention can detect an end point of a start of a speech and an end point of an end of a speech and judge it as a speech interval B, The section other than the section B is judged as the noise section (A, C).

In particular, the speech interval determination apparatus 100 according to an exemplary embodiment of the present invention performs a masking process and a normalization process according to an embodiment of the present invention to determine a speech interval while robustness against various noise environments.

In other words, in order to more clearly grasp the noise and the speech interval, the speech interval determination apparatus 100 of the present invention estimates an initial partial interval of the input speech signal as noise and outputs a speech characteristic corresponding to the estimated noise And performs a masking process of adding to the previous frame of the voice signal.

Thereafter, the speech interval determination apparatus 100 of the present invention extracts a speech feature vector on a frame-by-frame basis from the masked speech signal. Here, the speech feature vector that can be calculated may be, for example, energy information of a speech signal, pitch information, and the like. When energy information is used, Fourier transform is performed on the speech signal of the frame to output spectral information, And a process of outputting the log filter bank energy by applying a log function to the generated filter bank energy, thereby extracting the voice feature vector.

Hereinafter, the speech interval determination apparatus 100 of the present invention does not determine a speech interval using only one speech feature vector, but applies a sequence of speech feature vectors of a predetermined interval as an input value of a neural network model input layer , So that it is possible to detect the voice section.

The more specific configuration and operation of the speech interval determination apparatus 100 according to the present invention will be described later.

Meanwhile, the feature extracting apparatus 200 of the present invention extracts feature vectors for voice recognition corresponding to the voice interval detected through the voice interval determining apparatus 100. [ As described above, the feature vectors for speech recognition can utilize various speech features as well as energy information for the speech signal, and can be implemented by applying various feature extraction algorithms to the speech feature extraction. For example, the feature extracting apparatus 200 of the present invention may include a Mel-Frequency Cepstrum Coefficient (MFCC), a Linear Prediction Coefficient Cepstrum (LPCC), a Preceptual Linear Prediction Ceptrum Coeffcient (PLPCC), an EIH The feature parameters can be extracted by using various feature extraction algorithms such as Ensemble Interval Histogram (SMC) and Short-Time Modified Coherence (SMC).

Then, the feature extraction apparatus 200 transmits the extracted feature vector to the speech recognition apparatus 300.

The speech recognition apparatus 300 recognizes the feature vector extracted through the feature extraction apparatus 200 using the pre-built acoustic model 400 and outputs a speech recognition result corresponding thereto. In addition, the speech recognition apparatus 300 of the present invention can perform a speech recognition process by applying various speech recognition algorithms such as a Hidden Markov Model (HMM), a Dynamic Time Warping (DTW), and a neural network.

In addition, the acoustic model 400 of the present invention is constructed by statistically modeling phonemes. In the figure, only the acoustic model 400 is illustrated, but the acoustic model 400 is not limited to the acoustic model 400, A language model to be supported, a phonetic dictionary to support the calculation based on the standard phonetic method, and the like, and the final speech recognition result can be derived accordingly.

The result output through the feature extraction apparatus 200 and the result output through the speech recognition apparatus 300 may be reflected on the acoustic model 400 to construct the acoustic model 400 .

Hereinafter, the speech interval determination apparatus 100 of the present invention as described above will be described in more detail.

FIG. 3 is a block diagram for explaining a speech interval determination apparatus according to an embodiment of the present invention. FIGS. 4 to 6 are diagrams for explaining a speech interval determination method using normalization according to an embodiment of the present invention.

Referring to FIG. 3, the speech interval determination apparatus 100 of the present invention includes a masking processing unit 10, a feature vector extraction unit 20, a normalization unit 30, and a speech interval determination unit 40, . At this time, the speech interval determination apparatus 100 of the present invention may further include a frame formation unit (not shown) for separating the input speech signal into frames of a predetermined length. At this time, the frame forming unit (not shown) of the present invention can be separated into frames having a length of 20 to 30 msec every 10 msec.

Meanwhile, the masking processing unit 10 constituting the speech interval determination apparatus 100 of the present invention performs a preprocessing process for detecting a clearer noise from an input speech signal. As shown in FIG. 4, (A) in the speech signal. Then, the masking processing unit 10 adds the noise signal corresponding to the estimated noise period A to the entire voice signal. The reason that the masking processor 10 of the present invention estimates an initial part of a speech signal as a noise section is to add a speech energy of a lower level to a speech signal in comparison with speech characteristics of a speech section, The noise section is estimated on the assumption that an initial part of the speech signal is a noise section.

4, the masking processor 10 of the present invention adds the noise signal corresponding to the noise section A to the frame of the entire speech signal, and outputs the noise signal corresponding to the noise section A The size of the voice signal can be clarified, so that it is possible to more clearly judge the voice section and the voice section.

In addition, the masking processing unit 10 of the present invention can add a signal corresponding to an average energy level corresponding to the estimated noise period as a noise signal, and construct a learning DB 50 including learning data for voice interval determination The noise signal may be added to the noise signal. By performing such a masking process, it is possible to more clearly determine a speech interval even in a noise interval not considered in constructing learning data.

The masking process in the masking processing unit 10 can be calculated by the following equation.

는 입력 음성 신호이고,

는

를 마스킹하기 위한 잡음 신호이다.

와

의 초기 구간

샘플에서의 에너지 값의 제곱근(root mean square)

와

을 이용하여 입력 음성 신호의 초기 잡음 레벨의 수준으로

를 맞추어 마스킹을 수행하게 된다. here

Is an input speech signal,

The

As shown in FIG.

Wow

The initial section of

The root mean square of the energy value in the sample

Wow

To the level of the initial noise level of the input speech signal

And masking is performed.

에 대응하여 음성 특징 벡터를 추출하는 과정을 수행할 수 있다. The feature vector extracting unit 20 extracts the masked signal

And extracting speech feature vectors corresponding to the speech feature vectors.

In particular, the feature vector extraction unit 20 of the present invention can extract the vector value of the log mel-filter bank energy (LMFE) as a speech feature vector. For this, the feature vector extracting unit 20 of the present invention extracts spectral information by performing Fourier transform on the speech signal on a frame-by-frame basis, and calculates the filter bank energy for the spectrum. That is, the frequency band can be divided into several filter banks, and the logarithmic function can be applied to the energy in each bank to calculate the log filter bank energy. Accordingly, the feature vector extracting unit 20 of the present invention can calculate approximately 40-dimensional speech feature vectors for each frame.

The speech interval determination unit 40 can determine the speech interval using the speech feature vector extracted through the feature vector extraction unit 20. However, considering the real-time processing and various noise environments, the speech interval determination unit 40 can more accurately determine the speech interval The normalization unit 30 performs a normalization process of converting a speech feature vector from 0 to 1.

That is, the normalization unit 40 of the present invention performs a normalization process of converting the voice feature vector extracted through the feature vector extraction unit 20 from 0 to 1. At this time, the normalization unit 40 of the present invention performs a normalization process, The normalization process is performed considering the input layer of the learning model.

)를 입력 값으로 이용하게 된다. More specifically, the speech interval determination unit 40 of the present invention to be described later determines a speech interval using a neural network learning model, in particular, a neural network model. In this case, as shown in FIG. 5, the depth neural network model may include an input layer, a hidden layer, and an output layer, and derives an output value of the output layer based on input values input to the input layer. In this case, the speech interval determination unit 40 of the present invention not only considers only one speech feature vector extracted from a single frame as an input value of an input layer, but refers to a series of normalized speech features vector(

) As an input value.

Here, the series of speech feature vectors are determined in consideration of the size of the input layer of the input neural network model. The normalization unit 30 of the present invention divides the size of the input layer of the neural network model into a splice window, 600) method. That is, as shown in FIG. 6, the normalization unit 30 of the present invention normalizes speech feature vectors in splice window units. For example, the normalization unit 30 normalizes speech feature vectors in a certain range (m-5 to m + 5), and the normalization process can be performed within the range using the average and variance.

The normalization process according to the embodiment of the present invention can be defined according to the following equation.

와

는

번째 입력되는 음성 신호에서

번째 프레임의

번째 주파수 대역에 대한 음성 특징 벡터를 스플라이스 윈도우 단위로 추정한 평균과 표준편차를 의미하며,

은 스플라이스 윈도우 내의 전체 프레임수를 의미하는 것으로,

으로 나타낼 수 있다. 이때,

이고,

이며,

은 스플라이스 윈도우의 크기를 의미한다. here

Wow

The

Lt; th >

Th frame

(B) is a mean and standard deviation of a speech feature vector for a first frequency band in a splice window,

Is the total number of frames in the splice window,

. At this time,

ego,

Lt;

Is the size of the splice window.

번째 입력되는 음성 신호의

번째 프레임에 대한 스플라이스 윈도우에 속하는

번째 프레임 음성 특징 벡터의

번째 주파수 대역은 스플라이스 윈도우 내의 음성 특징 벡터를 대상으로 정규화 되는데,

가 정규화된 결과이다. As described above, the normalization unit 30 of the present invention

Of the input audio signal

Lt; RTI ID = 0.0 > frame < / RTI >

Th frame speech feature vector

Th frequency band is normalized to speech feature vectors in the splice window,

Is the normalized result.

번째 프레임의 스플라이스 윈도우 내에서 정규화된 음성 특징 벡터의 시퀀스

과 같이 구성되며,

로 발화에 대한 인덱스

은 생략했다.

은

번째 프레임에서 정규화된

차원의 음성 특징 벡터를 의미한다. 정규화된 음성 특징 벡터 시퀀스

은 심층 신경망의 입력 레이어의 입력 값으로 이용되게 된다.

A sequence of normalized speech feature vectors within the splice window of the < RTI ID = 0.0 >

Respectively,

Index for ignition as

Is omitted.

silver

Lt; RTI ID = 0.0 >

Dimensional speech feature vector. Normalized speech feature vector sequence

Is used as the input value of the input layer of the neural network.

As described above, the present invention realizes more real-time processing by performing normalization on a splice window basis considering an input value of a neural network model used for voice interval determination, It is possible to exert excellent performance.

The speech interval determination unit 40 determines whether or not the speech feature vector transmitted through the normalization unit 30 is speech by applying the speech feature vector to a neural network model, in particular, a deep neural network (DNN) model. The depth neural network model of the present invention includes an input layer and an output layer, and includes a multilayer perceptron (multi) layer including a hidden layer hidden between an input layer and an output layer. layer perceptron). Each layer may be composed of a plurality of nodes corresponding to the artificial neurons, and the learning relationship between the nodes of the different layers may be determined. In particular, the output value at a node is determined by the output value of the activation function of the node, and the input of the activation function may mean a weighted sum of all nodes connected to the node.

를 입력 레이어의 입력 값으로 적용하고, 출력 레이어에서는 m 번째 프레임에서의 음성 검출 결과(0 또는 1)이 출력되도록 적용한 후 신경망 알고리즘을 적용하여 음성 여부를 검출하게 된다. 5, the speech interval determiner 40 of the present invention determines whether speech is normal or not by using the depth neural network model. In particular, the speech interval determiner 40 of the present invention includes a normalizer 30, A sequence of normalized speech feature vectors

Is applied as an input value of the input layer, and the output layer is applied to output the voice detection result (0 or 1) in the m-th frame, and then the neural network algorithm is applied to detect the voice.

As described above, according to the present invention, a splice window is determined in consideration of a size of an input layer of a neural network model, an average and a variance of a speech feature vector calculated for each frequency band are calculated for each splice window, By performing the normalization on a window-by-window basis, it is possible to detect the presence or absence of speech more quickly. Also, before the extraction of the speech feature vector, the masking process is performed to extract the noise region more clearly considering various noise environments.

Hereinafter, a method of determining a voice interval using normalization according to an embodiment of the present invention will be described in more detail.

7 is a flowchart illustrating a method of determining a voice interval using normalization according to an embodiment of the present invention.

Referring to FIG. 7, the speech interval determination apparatus 100 of the present invention estimates an initial part of an input speech signal as a noise interval (S103) when a speech signal is input (S101) Noise signal to the entire speech signal. Here, the voice interval determination apparatus 100 of the present invention estimates an initial partial interval of a voice signal as a noise interval, by adding a voice characteristic of a voice interval, for example, And estimates a noise interval on the assumption that an initial part of the speech signal is a noise interval.

In addition, the speech interval determination apparatus 100 of the present invention performs a masking process of adding a noise signal corresponding to a noise interval to a frame of the entire speech signal (S105). Thereby, the size of the speech signal corresponding to the noise interval Can be made more clear, so that it becomes possible to more clearly judge the voice interval and the non-voice interval.

At this time, the speech interval determination apparatus 100 of the present invention may add a signal corresponding to the average energy level corresponding to the estimated noise period as a noise signal, and may add the noise signal applied at the time of building the learning DB 50 to the noise signal . By performing such a masking process, it is possible to more clearly determine the speech interval even in the noise interval not considered in the construction of the learning DB.

Thereafter, the speech interval determination apparatus 100 of the present invention extracts speech feature vectors using the masked speech signal (S107). For example, the speech interval determination apparatus 100 of the present invention can extract a speech feature vector from a vector value of a log mel-filter bank energy (LMFE) for a frame. To this end, the speech interval determination apparatus 100 of the present invention performs Fourier transform on a speech signal on a frame-by-frame basis, extracts spectrum information, and calculates a filter bank energy for the spectrum. That is, the frequency band can be divided into several filter banks, and the log filter bank energy can be calculated for each frequency band by applying the log function to the energy in each bank. Accordingly, the feature vector extracting unit 20 of the present invention can calculate approximately 40-dimensional speech feature vectors for each frame.

Subsequently, the speech interval determination apparatus 100 of the present invention normalizes the feature vectors on a splice window basis (S109).

For example, the speech interval determination apparatus 100 of the present invention calculates an average and a variance of a frame within a certain range (m-5 to m + 5) on the basis of the current frame m, The normalization process is performed. At this time, when calculating the average and variance, the average and variance of the frames can be calculated using the vector values of the log filter bank energy, which is a speech feature vector calculated for each frequency band in the corresponding frame, It is possible to perform a normalization process of setting 0 to 1 as a unit. In addition, the splice window, which is a normalization unit, takes the size of a neural network model input layer used for voice interval determination into consideration, performs normalization in units of the splice windows considered, By considering dispersion, it is possible to exert better performance in environmental distortion.

은 심층 신경망의 입력 레이어의 입력 값으로 설정하고, 출력 레이어의 출력 값으로 m 번째 프레임에서의 음성 검출 결과(0 또는 1)를 설정한 후 신경망 알고리즘을 적용하여 프레임 단위로 음성 여부를 검출할 수 있게 된다. The speech interval determination apparatus 100 of the present invention includes a normalized speech feature vector sequence

Is set to the input value of the input layer of the neural network, and the speech detection result (0 or 1) in the m-th frame is set as the output value of the output layer. .

The method of determining the speech interval using the normalization according to the embodiment of the present invention has been described above.

The method of determining a speech interval using the normalization of the present invention as described above may be provided in the form of a computer readable medium suitable for storing computer program instructions and data.

In particular, the computer program of the present invention is a speech segment determining apparatus for dividing a speech signal into frames having a predetermined length and detecting a speech segment by determining whether or not the segment is a speech segment, the apparatus comprising: Extracting a speech feature vector, performing a normalization on a series of the speech feature vectors according to an input layer of a deep neural network (DNN), and extracting the normalized speech feature vector from the depth- And a step of determining a voice interval in the voice signal by setting it as an input.

Such computer-readable media suitable for storing computer program instructions and data include, for example, magnetic media such as hard disks, floppy disks and magnetic tape, compact disk read only memory (CD-ROM) Optical media such as a DVD (Digital Video Disk), a magneto-optical medium such as a floppy disk, and a ROM (Read Only Memory), a RAM , Random Access Memory), flash memory, EPROM (Erasable Programmable ROM), and EEPROM (Electrically Erasable Programmable ROM). The processor and memory may be supplemented by, or incorporated in, special purpose logic circuits.

The computer readable recording medium may also be distributed over a networked computer system so that computer readable code can be stored and executed in a distributed manner. The functional program for implementing the present invention and the related code and code segment may be implemented by programmers in the technical field of the present invention in consideration of the system environment of the computer that reads the recording medium and executes the program, Or may be easily modified or modified by the user.

In addition, a computer program recorded on a computer-readable recording medium as described above includes instructions for performing the functions as described above, distributed and distributed through a recording medium, read and installed in a specific device, a specific computer, Thereby performing the above-described functions.

While the specification contains a number of specific implementation details, it should be understood that they are not to be construed as limitations on the scope of any invention or claim, but rather on the description of features that may be specific to a particular embodiment of a particular invention Should be understood. Certain features described herein in the context of separate embodiments may be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented in multiple embodiments, either individually or in any suitable subcombination. Further, although the features may operate in a particular combination and may be initially described as so claimed, one or more features from the claimed combination may in some cases be excluded from the combination, Or a variant of a subcombination.

Likewise, although the operations are depicted in the drawings in a particular order, it should be understood that such operations must be performed in that particular order or sequential order shown to achieve the desired result, or that all illustrated operations should be performed. In certain cases, multitasking and parallel processing may be advantageous. Also, the separation of the various system components of the above-described embodiments should not be understood as requiring such separation in all embodiments, and the described program components and systems will generally be integrated together into a single software product or packaged into multiple software products It should be understood.

The present invention relates to a method for determining a speech interval in an input speech signal, and more particularly, to a speech recognition method capable of real-time processing when determining a speech interval in a speech signal and at the same time, To a method for determining a speech interval using the normalization and a speech interval determination apparatus therefor.

According to the present invention, normalization is performed in consideration of the size of the input layer of the depth-of-field neural network model in determining the speech interval using the depth-of-field neural network model, so that the speech interval can be determined more quickly and accurately. The normalization process is performed after the speech recognition is performed, so that it is possible to determine a stronger speech interval in various noise environments. Accordingly, the present invention can improve the efficiency of the speech signal processing field such as speech recognition, speaker recognition, sound quality improvement, and the like, and thus is industrially applicable.

10:
20: Feature vector extraction unit
30: Normalization unit
40:
50: Learning DB
100: voice interval determination unit
200: Feature extraction device
300: Speech recognition device
400: acoustic model
500: Speech Recognition System

Claims (9)

1. A speech interval determining apparatus for classifying a speech signal into frames having a predetermined length and detecting a speech interval by determining whether or not the divided frames are speech,
Extracting a speech feature vector from an input speech signal;
Performing a normalization on a series of speech feature vectors according to an input layer of a deep neural network (DNN); And
Determining a speech interval in the speech signal by setting the normalized speech feature vector as an input to the neural network model;
And determining a voice interval using the normalization. The method according to claim 1,
Prior to extracting the speech feature vector,
Estimating a part of the input speech signal as a noise section;
Adding the noise signal corresponding to the estimated noise period to the entire voice signal and masking the noise signal;
Further comprising:
The step of extracting the speech feature vector
And extracting a speech feature vector from the masked speech signal. 3. The method of claim 2,
The noise signal
Wherein the training noise signal is a learning noise signal applied at the time of constructing learning data for the speech interval determination. The method of claim 3,
The masking step
Wherein the masking process is performed after adjusting the energy level of the learning noise signal to match the energy level of the noise signal. The method according to claim 1,
The step of performing the normalization comprises:
Setting a splice window in a certain range based on an arbitrary frame in the speech signal;
Calculating an average and a standard deviation corresponding to the splice window by using a speech feature vector extracted for each frequency band of the splice window; And
Performing normalization on the splice window using the calculated average and standard deviation;
And determining a voice interval using the normalization. 6. The method of claim 5,
Wherein the splice window is set corresponding to a size of an input layer of the depth-of-field neural network model. A computer-readable recording medium having recorded thereon a program for executing the method of determining a speech interval using the normalization according to any one of claims 1 to 6. 1. A speech interval determining apparatus for classifying a speech signal into frames having a predetermined length and detecting a speech interval by determining whether or not the divided frames are speech,
A feature vector extractor for extracting a speech feature vector from an input speech signal;
A normalization unit for performing normalization on a series of speech feature vectors extracted through the feature vector extraction unit according to an input layer of a deep neural network (DNN); And
A speech interval determiner for determining the speech interval in the speech signal by setting the normalized speech feature vector as an input to the depth neural network model;
Wherein the speech signal is a speech signal. 9. The method of claim 8,
A masking unit for estimating a part of the input speech signal as a noise section and adding a noise signal corresponding to the estimated noise section to the entire speech signal;
Further comprising: a speech processor;

Images